I just bought a house that I need to renovate, and there is nothing more hilarious than seeing a bunch of agents using architecture software at the speed of light. The future is so amazing. You don't need to know how to use any software. Agents will use it for you.

I increasingly think the way to predict who wins in consumer AI is to look at how much compute they have access to and what they pay for it. You can build an amazing product, invent a new form factor, find something users love. But software is becoming a commodity. Someone will copy your features, and then you’re competing on who can serve more users, at a lower cost. If they can serve 100m users while you’re stuck on a waitlist because you can’t get capacity, good luck. Especially if there are network effects. The companies that locked in gigantic amounts of compute 2–3 years ago have a huge advantage. Meanwhile you’re trying to buy capacity during a shortage and paying a premium for it. Same feature, same user, very different COGS. And if they train and serve their own models, they can optimize for their product and avoid paying someone else’s margin on every token. At scale that can make the gap even bigger. They can afford to give away something you need to charge for. All of this makes me think AI structurally favors incumbents. The capex they committed years ago becomes a cost and scale advantage across every consumer product they launch. They still need a good product and distribution. But if they get those right, competing with them is going to be very hard.

One of my more controversial bets: the AI feed will be the form factor of 2027! Most people don’t know what to prompt, and they won’t prompt their way into complicated agentic workflows. With chat, we ask users for digital labor: figure out what AI can do, explain the task, provide the context, repeat. That’s a lot to ask if we want billions of people using agents every day. Meanwhile, people already spend hours in feeds. Twitter, Instagram, LinkedIn. They know how to scroll, discover something useful, and act on it. An intelligent system that knows you should bring you the next useful action. Here’s the email you need to answer, already drafted. Here are the latest sales numbers, with an update ready to send to your team. You approve, edit, dismiss, or jump into chat to go deeper. That’s how people will discover what agents can do, without having to invent a prompt first. And you can put ads in a feed, one of the biggest business models on the internet. A system that understands your plans can recommend a relevant restaurant, product, or service, clearly labeled as sponsored. You get a reason to come back every day, useful work ready to approve, and a business model to fund the agents doing it.

Don’t be pessimistic. We’re about to see an avalanche of software. Custom software for every person, every workflow, every weird little problem that was never worth building for. I think we’ll have 1,000x more software from here. The end of generic software. The beginning of software for everyone.

The real SaaS apocalypse is pretty simple: every piece of software becomes reproducible. The cost of building it goes toward zero, competition becomes brutal, and eventually the price follows. Good luck charging for something anyone can recreate. Computer use is the breakthrough here imo. Here’s how it works: One agent navigates the existing software, clicks through every feature and documents everything in a markdown file. A swarm of agents builds it. Another agent acts as the verifier, opens the new software, clicks everywhere, checks that everything works, sends bugs back. Repeat. I think this scenario is very possible and very much in the work. But people are missing the other side which is the bull case for big tech cos. If you’re Adobe, you already have a massive test suite documenting how the software is supposed to behave, including all the backend stuff you can’t discover by clicking around. You have insane usage telemetry, decades of customer feedback, and you know which obscure features people actually depend on. You can build the same software factory. Except you start with a huge advantage. Use the agents to clean up decades of legacy code. Make the software load faster, feel faster, break less. Give it clean APIs that agents can use. Then use your distribution to ship that better product to millions of existing customers. Even if you burn $1M in tokens doing this, that’s roughly the annual cost of a handful of engineers in the Bay Area. And getting the mass market to switch takes time. Especially in the enterprise. A free open-source alternative still has to earn trust, build credibility and convince people to move their workflows. “We cloned Photoshop” doesn’t automatically do that. So there’s also a pretty strong bull case: the incumbents build the software factory first, use their tests, customer knowledge and distribution to move faster, and end up with better software, higher margins, more revenue and happier customers. If they don’t move fast enough, well, we’re back to the bear case.

AI is destroying the economic value of exclusivity. You can still own your code, your novel, your design. But if anyone can generate something equally good for free, what exactly are you charging for? You own something that is almost free

Having spent 10 years building AI products, this resonates a lot. Machine learning has long been a black box. Now we’re giving a much more capable black box access to our most sensitive data and the ability to act on our behalf. Alignment is still a huge challenge. We may never fully understand why a model made a particular decision or which training data shaped it. We need guardrails around it that hold even when the model gets things wrong. An email agent can have read-only access: it can summarize your inbox, but it has no permission to send anything. Through Entra ID, an agent acting on your behalf stays within explicitly granted permissions and your existing access. These boundaries need to be enforced outside the model. Same for action logs, spending limits and the ability to stop it mid-task. “Don’t do this” in a prompt is a request. Removing the permission is a control.

This is everything I admired about America as a teenager growing up in France. Michael Dell starts a company at 19 in his dorm room with $1,000, becomes a billionaire, then he and Susan commit $6.25 billion to give 25 million American children their first investment. I remember my high school teachers blaming large corporations and their dividends. It was just accepted in the classroom: corporations were evil, dividends were evil. And I remember thinking: wait. There’s a system where I can buy a piece of these companies, and they’ll give me money on a recurring basis? Everyone was explaining why this was terrible, and all I could think was: I need a piece of that. If I want to escape poverty, I need to own some of these businesses. It took me years to understand investing properly. I hope these kids get that realization much sooner. Munger said: “Show me the incentive and I’ll show you the outcome.” You can explain the benefits of capitalism forever. But owning a piece of a business gives you a personal reason to understand profits, dividends, reinvestment and compounding. Suddenly, America’s most successful companies are partly yours. Their success means something to you. And maybe a kid who starts by asking “What do I own?” will eventually ask “What could I build?” That’s what makes this so inspiring to me. A man who built his fortune giving millions of children a first experience of ownership, and perhaps the ambition to build something themselves. This is the America I dreamed of. LFG America 🇺🇸

The most successful software franchise of all time is coming to chat. Almost every knowledge worker on the planet uses Microsoft Office. What’s fascinating is how enduring this franchise has been. Born in the 1980s, it survived every major platform shift: desktop, web, mobile, Windows, Mac, iOS, Android etc! And now, 40 years later, Word, Excel and PowerPoint are becoming AI-native inside Copilot Chat. Just the next evolution of one of the most enduring franchises in software history. It’s beautiful!

I think the people who spent their careers betting on quantum computing are going to look very smart. They spent decades building the hardware, the labs and the expertise, and now imagine putting AI that’s superhuman at math and physics on top of all that infrastructure. You could run thousands of experiments, use the results to improve both the hardware and the software, then repeat, getting closer to reliable, error-corrected quantum computing with every cycle. And once that works, you can turn it toward simulating molecules, discovering new materials and understanding the chemistry behind biology. My bet is that AI will massively accelerate what these people spent their lives building.

Feels inevitable to me. If AI keeps getting better at math, it could discover a shortcut for factoring huge numbers, a new attack on elliptic curves, or an entirely different way around the assumptions behind encryption. Every major lab will race toward this because the advantage would be enormous, both commercially and for national security. You want to be the first to find the weakness and secure your own infrastructure when trillions of dollars flow through it. Btw, Matthew Green is a cryptography professor at Johns Hopkins. He also co-created Zerocash, the privacy protocol behind Zcash, so this is very much his field.

Some people are anti-AI because they can already see the downside, potentially losing their job or watching the craft they spent years learning get automated, while the promised land, curing cancer and treating diseases, still feels very far away. AI gets incredibly useful when it has good data and a way to test whether it’s right. But biology has an enormous space of possibilities, much of the relevant data hasn’t been generated yet, and you can’t answer every question by thinking harder. At some point you have to run the experiment. So here’s something I’d love to see someone build: a giant warehouse, basically a data center for biology, filled with lab equipment and eventually tens of thousands of robots running experiments around the clock. Liquid handlers, plate readers, microscopes, incubators, all connected to models that propose experiments, look at the results and decide what to try next. I call the idea bounded serendipity. A lot of discovery comes from exploring something and finding something you weren’t looking for. CRISPR came out of research into how bacteria defend themselves against viruses. RNA interference also emerged from unexpected experimental observations. These were years of careful science, but the surprise mattered. You can’t schedule the breakthrough, you can create more opportunities to encounter one. So why wouldn’t we massively scale those opportunities? Pick a scientifically meaningful area, define what can be tested and measured reliably, then explore far more of it than a human team could. Give the models room to try unusual ideas, follow surprising results and update their understanding, with scientists deciding which questions matter and checking what the results actually mean. Two things make me think the economics could change. 1/ Robotics. Lab automation already exists, but integrating different instruments and workflows often takes expensive custom engineering. My bet is that increasingly capable general-purpose robots, including humanoids where they make sense, could reduce how much bespoke hardware you need by using equipment designed for humans. Keep specialized machines for the tasks they do best, and use flexible robots to connect everything and handle the work between them. We’re not yet at the point where you can buy 10,000 humanoids and expect them to run a biology lab reliably. But that’s the direction I’d bet on. Once a workflow works, you can replicate it and run shifts around the clock. Scaling that capacity could become much easier than recruiting and training thousands of lab technicians for every new facility. 2/ The data itself can be a product. Historically, the economics of an experiment often depended on whether it helped produce a valuable discovery. Now there’s another potential buyer: someone training a biology model. A well-run experiment that shows something doesn’t work can still be useful training data. A broken experiment is different, but a clean negative result tells you something about the world. That creates a possible business model where you sell high-quality experimental datasets while also using them to build better models and pursue your own discoveries. Companies are already selling biological data for AI training, so this part is more than a hypothetical. And parts of the loop already work. OpenAI and Ginkgo connected GPT-5 to an automated lab, tested more than 36,000 protein-synthesis reaction compositions and reported a 40% reduction in production cost. That’s a narrow result, but it’s a concrete example of a model proposing experiments, getting real measurements back and improving through iteration. The flywheel is more experiments, more useful data, better models, smarter experiments, and data revenue funding more capacity. We’re building enormous data centers for compute. I’d love to see us build the equivalent for biology, so the promised land gets here faster. Bounded serendipity at industrial scale.

Looking at OpenAI releasing hundreds of mathematical results, I think we really have to lead with empathy here. When you’ve spent years learning a craft, whether it’s writing code, creating illustrations or doing mathematics, and suddenly a lot of what you do seems to be a prompt away, of course that’s disorienting. Your craft is how you define yourself, what you’re proud of, where you find flow, sometimes the thing that gives your day meaning. It’s a huge part of your identity. And I get how this can lead to a pretty nihilistic feeling, like you wake up one morning and think, okay, if everything can just be prompted into existence, what am I supposed to do now? What am I getting better at? What is all this effort for? Software engineers have been grappling with this for a while, and I expect more people will experience it as AI keeps accelerating. Writing a book, creating a movie, maybe someday helping solve a disease, more and more things we’ve spent our lives learning could feel a prompt away. I think that progress is amazing, and I’m 100% confident we’ll find harder problems to work on and new crafts to fall in love with. But knowing that intellectually doesn’t make the sense of loss disappear, especially when the change happens this fast and feels this brutal. We can celebrate the breakthroughs and still care about the millions of people trying to figure out where they fit and what gives their life meaning in a world becoming this intelligent. “You can do more now” might be true, but people also need time to process what they feel they’ve lost.